Why Jig?

Every agent framework claims safety. But they all stop at prompt-level guardrails — suggestions, not enforcement. Jig is different.

Jig intercepts every tool call at the code level, before execution. Not in a prompt. Not after the fact. Before.

# deepagents: "You should not delete files..." — Agent ignores this
# Jig:
from jig.adapters.mcp_client import ToolGuard
ToolGuard.check("coding", "Bash(rm -rf /)")  # → False, blocked at code level

While other frameworks "trust the LLM" (deepagents' own words), Jig enforces boundaries that agents cannot bypass.

Quick Start

pip install jig-toolguard

from jig import Jig

# Load 11 built-in agents and run a pipeline
app = Jig(skills_dir="./skills")
result = app.run("Review this code for security issues")
print(result)

# Clone and run
git clone https://github.com/luyi14-bits/jig.git
cd jig
pip install -e .
echo "exit" | python run.py

Install

pip install jig-toolguard

Or from source:

git clone https://github.com/luyi14-bits/jig.git
cd jig
pip install -e .

Requires Python 3.10+. Set your DeepSeek API key:

export DEEPSEEK_API_KEY="sk-your-key"

Features

🛡️ Pre-execution Safety (Agent Firewall)

Every tool call passes through three layers before execution:

1. Whitelist — agent role must have explicit permission for the tool
2. Denylist — dangerous operations (rm -rf /, drop table) globally blocked
3. Runtime hook — custom validation callback

No other agent framework implements code-level pre-execution interception.

🧠 DeepSeek-Native Optimization

• SHA-256 prefix caching (session-level byte invariance)
• Flash-first cost-aware routing (auto-upgrade to Pro for complex tasks)
• Automatic Function Calling repair

🤖 11 Preset Agents

🌐 Multi-Model + Streaming

• DeepSeekProvider — fully optimized with caching and FC repair
• OpenAIProvider — extensible interface
• SSE streaming — POST /stream endpoint with real-time events

🔄 Loop Engineering

• Convergence detection — automatic stop when output stabilizes
• Quality validation — per-node output quality scoring
• Checkpoint resume — crash recovery without re-running completed nodes
• CircuitBreaker — 3-state (CLOSED/OPEN/HALF_OPEN) fault isolation

🗄️ 4-Layer Memory

🧩 Graph Orchestration

• DAG-based pipeline with conditional routing, parallel branches, and self-loops
• Replace linear SOPNode with dynamic GraphOrchestrator

Comparison

Architecture

Agent Firewall (Control Plane):
  ToolGuard · LOOP SOP · CircuitBreaker · GlobalConstraints

Agent Plane:
  SkillParser → SkillRegistry → AgentFactory → Agents (SKILL.md)

Orchestration Plane:
  SOPRunner · GraphOrchestrator · LoopEngine · Memory · Checkpoint

Tool Plane:
  MCP · ModelRouter · CacheEngine · CostAwareRouter · Streaming

Project Structure

├── src/jig/           # Framework core (adapters/core/orchestrator/cli/server)
├── tests/             # 124 test suite
├── skills/            # 11 Agent SKILL.md definitions
├── docs/              # Guides · Reports · Whitepapers · Blog
│   ├── guides/        # User guides
│   ├── reports/       # Comparison · Gap analysis · Security audits
│   └── whitepapers/   # Technical whitepapers v1–v4
├── examples/          # Example projects
├── scripts/           # Utility scripts
├── .github/           # Issue templates · PR template
├── pyproject.toml     # Build config → pip install jig-toolguard
└── CHANGELOG.md

Roadmap

Contributing

See CONTRIBUTING.md. PRs welcome!

License

MIT — Copyright (c) 2026 Jig Contributors.